Human Action Recognition Using Supervised pLSA

Probabilistic latent semantic analysis (pLSA) has been widely used by researchers for human action recognition from video sequences. However, one of the major disadvantages of pLSA and its other extensions is that category labels of training samples are not fully used in model learning procedure for classification task. In this paper, a supervised pLSA (spLSA) model is proposed for overcoming this drawback. By adding an observable category variable to generative process of classic pLSA, spLSA is endowed with more discriminative power. Thus, this model provides a unified framework for semantic analysis and object classification, where the topics formulation is guided by spLSA towards more discriminative and the mapping between the topics and the action categories are described in a fully probabilistic manner. Experimental results show that spLSA substantially outperforms pLSA and achieves comparable or better performances than latent dirichlet allocation based supervised models and other state-of-the-art methods.